[ceph.git] / ceph / src / spdk / test / vhost / perf_bench / vhost_perf.sh

#!/usr/bin/env bash

testdir=$(readlink -f $(dirname $0))
rootdir=$(readlink -f $testdir/../../..)
source $rootdir/test/common/autotest_common.sh
source $rootdir/test/vhost/common.sh

vhost_num="0"
vm_memory=2048
vm_sar_enable=false
host_sar_enable=false
sar_delay="0"
sar_interval="1"
sar_count="10"
vm_throttle=""
ctrl_type="spdk_vhost_scsi"
use_split=false
kernel_cpus=""
run_precondition=false
lvol_stores=()
lvol_bdevs=()
split_bdevs=()
used_vms=""
wwpn_prefix="naa.5001405bc6498"
packed_ring=false

fio_iterations=1
fio_gtod=""
precond_fio_bin=$CONFIG_FIO_SOURCE_DIR/fio
disk_map=""

disk_cfg_bdfs=()
disk_cfg_spdk_names=()
disk_cfg_splits=()
disk_cfg_vms=()
disk_cfg_kernel_names=()

function usage() {
	[[ -n $2 ]] && (
		echo "$2"
		echo ""
	)
	echo "Shortcut script for doing automated test"
	echo "Usage: $(basename $1) [OPTIONS]"
	echo
	echo "-h, --help                  Print help and exit"
	echo "    --fio-bin=PATH          Path to FIO binary on host.;"
	echo "                            Binary will be copied to VM, static compilation"
	echo "                            of binary is recommended."
	echo "    --fio-jobs=PATH         Comma separated list of fio config files to use for test."
	echo "    --fio-iterations=INT    Number of times to run specified workload."
	echo "    --fio-gtod-reduce       Enable fio gtod_reduce option in test."
	echo "    --vm-memory=INT         Amount of RAM memory (in MB) to pass to a single VM."
	echo "                            Default: 2048 MB"
	echo "    --vm-image=PATH         OS image to use for running the VMs."
	echo "                            Default: \$HOME/vhost_vm_image.qcow2"
	echo "    --vm-sar-enable         Measure CPU utilization in guest VMs using sar."
	echo "    --host-sar-enable       Measure CPU utilization on host using sar."
	echo "    --sar-delay=INT         Wait for X seconds before starting SAR measurement. Default: 0."
	echo "    --sar-interval=INT      Interval (seconds) argument for SAR. Default: 1s."
	echo "    --sar-count=INT         Count argument for SAR. Default: 10."
	echo "    --vm-throttle-iops=INT  I/Os throttle rate in IOPS for each device on the VMs."
	echo "    --ctrl-type=TYPE        Controller type to use for test:"
	echo "                            spdk_vhost_scsi - use spdk vhost scsi"
	echo "                            spdk_vhost_blk - use spdk vhost block"
	echo "                            kernel_vhost - use kernel vhost scsi"
	echo "                            Default: spdk_vhost_scsi"
	echo "    --packed-ring           Use packed ring support. Requires Qemu 4.2.0 or greater. Default: disabled."
	echo "    --use-split             Use split vbdevs instead of Logical Volumes"
	echo "    --limit-kernel-vhost=INT  Limit kernel vhost to run only on a number of CPU cores."
	echo "    --run-precondition      Precondition lvols after creating. Default: true."
	echo "    --precond-fio-bin       FIO binary used for SPDK fio plugin precondition. Default: $CONFIG_FIO_SOURCE_DIR/fio."
	echo "    --custom-cpu-cfg=PATH   Custom CPU config for test."
	echo "                            Default: spdk/test/vhost/common/autotest.config"
	echo "    --disk-map              Disk map for given test. Specify which disks to use, their SPDK name,"
	echo "                            how many times to split them and which VMs should be attached to created bdevs."
	echo "                            Example:"
	echo "                            NVME PCI BDF,Spdk Bdev Name,Split Count,VM List"
	echo "                            0000:1a:00.0,Nvme0,2,0 1"
	echo "                            0000:1b:00.0,Nvme1,2,2 3"
	echo "-x                          set -x for script debug"
	exit 0
}

function cleanup_lvol_cfg() {
	notice "Removing lvol bdevs"
	for lvol_bdev in "${lvol_bdevs[@]}"; do
		$rpc_py bdev_lvol_delete $lvol_bdev
		notice "lvol bdev $lvol_bdev removed"
	done

	notice "Removing lvol stores"
	for lvol_store in "${lvol_stores[@]}"; do
		$rpc_py bdev_lvol_delete_lvstore -u $lvol_store
		notice "lvol store $lvol_store removed"
	done
}

function cleanup_split_cfg() {
	notice "Removing split vbdevs"
	for disk in "${disk_cfg_spdk_names[@]}"; do
		$rpc_py bdev_split_delete ${disk}n1
	done
}

function cleanup_parted_config() {
	notice "Removing parted disk configuration"
	for disk in "${disk_cfg_kernel_names[@]}"; do
		parted -s /dev/${disk}n1 rm 1
	done
}

function cleanup_kernel_vhost() {
	notice "Cleaning kernel vhost configration"
	targetcli clearconfig confirm=True
	cleanup_parted_config
}

function create_vm() {
	vm_num=$1
	setup_cmd="vm_setup --disk-type=$ctrl_type --force=$vm_num --memory=$vm_memory --os=$VM_IMAGE"
	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
		x=$(printf %03d $vm_num)
		setup_cmd+=" --disks=${wwpn_prefix}${x}"
	else
		setup_cmd+=" --disks=0"
	fi

	if $packed_ring; then
		setup_cmd+=" --packed"
	fi

	$setup_cmd
	used_vms+=" $vm_num"
	echo "Added to used vms"
	echo $used_vms
}

function create_spdk_controller() {
	vm_num=$1
	bdev=$2

	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
		$rpc_py vhost_create_scsi_controller naa.0.$vm_num
		notice "Created vhost scsi controller naa.0.$vm_num"
		$rpc_py vhost_scsi_controller_add_target naa.0.$vm_num 0 $bdev
		notice "Added LUN 0/$bdev to controller naa.0.$vm_num"
	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
		if $packed_ring; then
			p_opt="-p"
		fi

		$rpc_py vhost_create_blk_controller naa.0.$vm_num $bdev $p_opt
		notice "Created vhost blk controller naa.0.$vm_num $bdev"
	fi
}

while getopts 'xh-:' optchar; do
	case "$optchar" in
		-)
			case "$OPTARG" in
				help) usage $0 ;;
				fio-bin=*) fio_bin="--fio-bin=${OPTARG#*=}" ;;
				fio-jobs=*) fio_jobs="${OPTARG#*=}" ;;
				fio-iterations=*) fio_iterations="${OPTARG#*=}" ;;
				fio-gtod-reduce) fio_gtod="--gtod-reduce" ;;
				vm-memory=*) vm_memory="${OPTARG#*=}" ;;
				vm-image=*) VM_IMAGE="${OPTARG#*=}" ;;
				vm-sar-enable) vm_sar_enable=true ;;
				host-sar-enable) host_sar_enable=true ;;
				sar-delay=*) sar_delay="${OPTARG#*=}" ;;
				sar-interval=*) sar_interval="${OPTARG#*=}" ;;
				sar-count=*) sar_count="${OPTARG#*=}" ;;
				vm-throttle-iops=*) vm_throttle="${OPTARG#*=}" ;;
				ctrl-type=*) ctrl_type="${OPTARG#*=}" ;;
				packed-ring) packed_ring=true ;;
				use-split) use_split=true ;;
				run-precondition) run_precondition=true ;;
				precond-fio-bin=*) precond_fio_bin="${OPTARG#*=}" ;;
				limit-kernel-vhost=*) kernel_cpus="${OPTARG#*=}" ;;
				custom-cpu-cfg=*) custom_cpu_cfg="${OPTARG#*=}" ;;
				disk-map=*) disk_map="${OPTARG#*=}" ;;
				*) usage $0 "Invalid argument '$OPTARG'" ;;
			esac
			;;
		h) usage $0 ;;
		x)
			set -x
			x="-x"
			;;
		*) usage $0 "Invalid argument '$OPTARG'" ;;
	esac
done

rpc_py="$rootdir/scripts/rpc.py -s $(get_vhost_dir 0)/rpc.sock"

if [[ -n $custom_cpu_cfg ]]; then
	source $custom_cpu_cfg
	vhost_reactor_mask="vhost_${vhost_num}_reactor_mask"
	vhost_reactor_mask="${!vhost_reactor_mask}"
	vhost_master_core="vhost_${vhost_num}_master_core"
	vhost_master_core="${!vhost_master_core}"
fi

if [[ -z $fio_jobs ]]; then
	error "No FIO job specified!"
fi

trap 'error_exit "${FUNCNAME}" "${LINENO}"' INT ERR

if [[ -z $disk_map ]]; then
	fail "No disk map provided for test. Exiting."
fi

# ===== Precondition NVMes if specified =====
if [[ $run_precondition == true ]]; then
	# Using the same precondition routine possible for lvols thanks
	# to --clear-method option. Lvols should not UNMAP on creation.
	json_cfg=$rootdir/nvme.json
	cat <<- JSON > "$json_cfg"
		{"subsystems":[
		 $("$rootdir/scripts/gen_nvme.sh" --json)
		]}
	JSON
	mapfile -t nvmes < <(grep -oP "Nvme\d+" "$json_cfg")
	fio_filename=$(printf ":%sn1" "${nvmes[@]}")
	fio_filename=${fio_filename:1}
	$precond_fio_bin --name="precondition" \
		--ioengine="${rootdir}/build/fio/spdk_bdev" \
		--rw="write" --spdk_json_conf="$json_cfg" --thread="1" \
		--group_reporting --direct="1" --size="100%" --loops="2" --bs="256k" \
		--iodepth=32 --filename="${fio_filename}" || true
fi

set +x
readarray disk_cfg < $disk_map
for line in "${disk_cfg[@]}"; do
	echo $line
	IFS=","
	s=($line)
	disk_cfg_bdfs+=(${s[0]})
	disk_cfg_spdk_names+=(${s[1]})
	disk_cfg_splits+=(${s[2]})
	disk_cfg_vms+=("${s[3]}")

	# Find kernel nvme names
	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
		tmp=$(find /sys/devices/pci* -name ${s[0]} -print0 | xargs sh -c 'ls $0/nvme')
		disk_cfg_kernel_names+=($tmp)
		IFS=" "
	fi
done
unset IFS
set -x

if [[ "$ctrl_type" == "kernel_vhost" ]]; then
	notice "Configuring kernel vhost..."
	trap 'vm_kill_all; sleep 1; cleanup_kernel_vhost; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR

	# Split disks using parted for kernel vhost
	newline=$'\n'
	backstores=()
	for ((i = 0; i < ${#disk_cfg_kernel_names[@]}; i++)); do
		nvme=${disk_cfg_kernel_names[$i]}
		splits=${disk_cfg_splits[$i]}
		notice "  Creating extended partition on disk /dev/${nvme}n1"
		parted -s /dev/${nvme}n1 mklabel msdos
		parted -s /dev/${nvme}n1 mkpart extended 2048s 100%

		part_size=$((100 / ${disk_cfg_splits[$i]})) # Split 100% of disk into roughly even parts
		echo "  Creating  ${splits} partitions of relative disk size ${part_size}"
		for p in $(seq 0 $((splits - 1))); do
			p_start=$((p * part_size))
			p_end=$((p_start + part_size))
			parted -s /dev/${nvme}n1 mkpart logical ${p_start}% ${p_end}%
			sleep 3
		done

		# Prepare kernel vhost configuration
		# Below grep: match only NVMe partitions which are not "Extended" type.
		# For example: will match nvme0n1p15 but not nvme0n1p1
		partitions=$(find /dev -name "${nvme}n1*" | sort --version-sort | grep -P 'p(?!1$)\d+')
		# Create block backstores for vhost kernel process
		for p in $partitions; do
			backstore_name=$(basename $p)
			backstores+=("$backstore_name")
			targetcli backstores/block create $backstore_name $p
		done
		partitions=($partitions)

		# Create kernel vhost controllers and add LUNs
		# Setup VM configurations
		vms_to_run=(${disk_cfg_vms[i]})
		for ((j = 0; j < ${#vms_to_run[@]}; j++)); do
			# WWPN prefix misses 3 characters. Need to complete it
			# using block backstore number
			x=$(printf %03d ${vms_to_run[$j]})
			wwpn="${wwpn_prefix}${x}"
			targetcli vhost/ create $wwpn
			targetcli vhost/$wwpn/tpg1/luns create /backstores/block/$(basename ${partitions[$j]})
			create_vm ${vms_to_run[j]}
			sleep 1
		done
	done
	targetcli ls
else
	notice "Configuring SPDK vhost..."
	vhost_run "${vhost_num}" "--no-gen-nvme" "-p ${vhost_master_core}" "-m ${vhost_reactor_mask}"
	notice "..."

	if [[ $use_split == true ]]; then
		notice "Configuring split bdevs configuration..."
		trap 'cleanup_split_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
		for ((i = 0; i < ${#disk_cfg_bdfs[@]}; i++)); do
			nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
			notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"

			splits=$($rpc_py bdev_split_create $nvme_bdev ${disk_cfg_splits[$i]})
			splits=($splits)
			notice "Created splits: ${splits[*]} on Bdev ${nvme_bdev}"
			for s in "${splits[@]}"; do
				split_bdevs+=($s)
			done

			vms_to_run=(${disk_cfg_vms[i]})
			for ((j = 0; j < ${#vms_to_run[@]}; j++)); do
				notice "Setting up VM ${vms_to_run[j]}"
				create_spdk_controller "${vms_to_run[j]}" ${splits[j]}
				create_vm ${vms_to_run[j]}
			done
			echo " "
		done
		bdevs=("${split_bdevs[@]}")
	else
		notice "Configuring LVOLs..."
		trap 'cleanup_lvol_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
		for ((i = 0; i < ${#disk_cfg_bdfs[@]}; i++)); do
			nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
			notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"

			ls_guid=$($rpc_py bdev_lvol_create_lvstore $nvme_bdev lvs_$i --clear-method none)
			lvol_stores+=("$ls_guid")
			notice "Created Lvol Store: $ls_guid on Bdev $nvme_bdev"

			vms_to_run=(${disk_cfg_vms[i]})
			for ((j = 0; j < ${disk_cfg_splits[$i]}; j++)); do
				free_mb=$(get_lvs_free_mb "$ls_guid")
				size=$((free_mb / ((${disk_cfg_splits[$i]} - j))))
				lb_name=$($rpc_py bdev_lvol_create -u $ls_guid lbd_$j $size --clear-method none)
				lvol_bdevs+=("$lb_name")
				notice "Created LVOL Bdev $lb_name on Lvol Store $ls_guid on Bdev $nvme_bdev"

				notice "Setting up VM ${vms_to_run[j]}"
				create_spdk_controller "${vms_to_run[j]}" ${lb_name}
				create_vm ${vms_to_run[j]}
			done
			echo " "
		done
		$rpc_py bdev_lvol_get_lvstores
	fi
	$rpc_py bdev_get_bdevs
	$rpc_py vhost_get_controllers
fi

# Start VMs
# Run VMs
vm_run $used_vms
vm_wait_for_boot 300 $used_vms

if [[ -n "$kernel_cpus" ]]; then
	mkdir -p /sys/fs/cgroup/cpuset/spdk
	kernel_mask=$vhost_0_reactor_mask
	kernel_mask=${kernel_mask#"["}
	kernel_mask=${kernel_mask%"]"}

	echo "$kernel_mask" >> /sys/fs/cgroup/cpuset/spdk/cpuset.cpus
	echo "0-1" >> /sys/fs/cgroup/cpuset/spdk/cpuset.mems

	kernel_vhost_pids=$(pgrep "vhost" -U root)
	for kpid in $kernel_vhost_pids; do
		echo "Limiting kernel vhost pid ${kpid}"
		echo "${kpid}" >> /sys/fs/cgroup/cpuset/spdk/tasks
	done
fi

# Run FIO
fio_disks=""
for vm_num in $used_vms; do
	host_name="VM-$vm_num"
	vm_exec $vm_num "hostname $host_name"
	vm_start_fio_server $fio_bin $vm_num

	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
		vm_check_scsi_location $vm_num
	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
		vm_check_blk_location $vm_num
	elif [[ "$ctrl_type" == "kernel_vhost" ]]; then
		vm_check_scsi_location $vm_num
	fi

	if [[ -n "$vm_throttle" ]]; then
		block=$(printf '%s' $SCSI_DISK)
		major_minor=$(vm_exec "$vm_num" "cat /sys/block/$block/dev")
		vm_exec "$vm_num" "echo \"$major_minor $vm_throttle\" > /sys/fs/cgroup/blkio/blkio.throttle.read_iops_device"
		vm_exec "$vm_num" "echo \"$major_minor $vm_throttle\" > /sys/fs/cgroup/blkio/blkio.throttle.write_iops_device"
	fi

	fio_disks+=" --vm=${vm_num}$(printf ':/dev/%s' $SCSI_DISK)"
done

# Run FIO traffic
for fio_job in ${fio_jobs//,/ }; do
	fio_job_fname=$(basename $fio_job)
	fio_log_fname="${fio_job_fname%%.*}.log"
	for i in $(seq 1 $fio_iterations); do
		echo "Running FIO iteration $i for $fio_job_fname"
		run_fio $fio_bin --hide-results --job-file="$fio_job" --out="$VHOST_DIR/fio_results" --json $fio_disks $fio_gtod &
		fio_pid=$!

		if $host_sar_enable || $vm_sar_enable; then
			pids=""
			mkdir -p $VHOST_DIR/fio_results/sar_stats
			sleep $sar_delay
		fi

		if $host_sar_enable; then
			sar -P ALL $sar_interval $sar_count > "$VHOST_DIR/fio_results/sar_stats/sar_stats_host.txt" &
			pids+=" $!"
		fi

		if $vm_sar_enable; then
			for vm_num in $used_vms; do
				vm_exec "$vm_num" "mkdir -p /root/sar; sar -P ALL $sar_interval $sar_count >> /root/sar/sar_stats_VM${vm_num}_run${i}.txt" &
				pids+=" $!"
			done
		fi

		for j in $pids; do
			wait $j
		done

		if $vm_sar_enable; then
			for vm_num in $used_vms; do
				vm_scp "$vm_num" "root@127.0.0.1:/root/sar/sar_stats_VM${vm_num}_run${i}.txt" "$VHOST_DIR/fio_results/sar_stats"
			done
		fi

		wait $fio_pid
		mv $VHOST_DIR/fio_results/$fio_log_fname $VHOST_DIR/fio_results/$fio_log_fname.$i
		sleep 1
	done

	parse_fio_results "$VHOST_DIR/fio_results" "$fio_log_fname"
done

notice "Shutting down virtual machines..."
vm_shutdown_all

if [[ "$ctrl_type" == "kernel_vhost" ]]; then
	cleanup_kernel_vhost || true
else
	notice "Shutting down SPDK vhost app..."
	if [[ $use_split == true ]]; then
		cleanup_split_cfg
	else
		cleanup_lvol_cfg
	fi
	vhost_kill "${vhost_num}"
fi

if [[ -n "$kernel_cpus" ]]; then
	rmdir /sys/fs/cgroup/cpuset/spdk
fi
Commit	Line	Data
11fdf7f2	1	#!/usr/bin/env bash
11fdf7f2	2
9f95a23c TL	3	testdir=$(readlink -f $(dirname $0))
	4	rootdir=$(readlink -f $testdir/../../..)
	5	source $rootdir/test/common/autotest_common.sh
	6	source $rootdir/test/vhost/common.sh
	7
f67539c2	8	vhost_num="0"
11fdf7f2	9	vm_memory=2048
9f95a23c	10	vm_sar_enable=false
f67539c2 TL	11	host_sar_enable=false
	12	sar_delay="0"
	13	sar_interval="1"
	14	sar_count="10"
9f95a23c	15	vm_throttle=""
11fdf7f2 TL	16	ctrl_type="spdk_vhost_scsi"
11fdf7f2 TL	17	use_split=false
9f95a23c TL	18	kernel_cpus=""
9f95a23c TL	19	run_precondition=false
11fdf7f2 TL	20	lvol_stores=()
11fdf7f2 TL	21	lvol_bdevs=()
f67539c2	22	split_bdevs=()
11fdf7f2	23	used_vms=""
9f95a23c	24	wwpn_prefix="naa.5001405bc6498"
f67539c2	25	packed_ring=false
11fdf7f2	26
9f95a23c	27	fio_iterations=1
f67539c2 TL	28	fio_gtod=""
	29	precond_fio_bin=$CONFIG_FIO_SOURCE_DIR/fio
	30	disk_map=""
	31
	32	disk_cfg_bdfs=()
	33	disk_cfg_spdk_names=()
	34	disk_cfg_splits=()
	35	disk_cfg_vms=()
	36	disk_cfg_kernel_names=()
	37
	38	function usage() {
	39	[[ -n $2 ]] && (
	40	echo "$2"
	41	echo ""
	42	)
11fdf7f2 TL	43	echo "Shortcut script for doing automated test"
	44	echo "Usage: $(basename $1) [OPTIONS]"
	45	echo
	46	echo "-h, --help Print help and exit"
	47	echo " --fio-bin=PATH Path to FIO binary on host.;"
	48	echo " Binary will be copied to VM, static compilation"
	49	echo " of binary is recommended."
f67539c2	50	echo " --fio-jobs=PATH Comma separated list of fio config files to use for test."
9f95a23c	51	echo " --fio-iterations=INT Number of times to run specified workload."
f67539c2	52	echo " --fio-gtod-reduce Enable fio gtod_reduce option in test."
11fdf7f2 TL	53	echo " --vm-memory=INT Amount of RAM memory (in MB) to pass to a single VM."
	54	echo " Default: 2048 MB"
	55	echo " --vm-image=PATH OS image to use for running the VMs."
f67539c2 TL	56	echo " Default: \$HOME/vhost_vm_image.qcow2"
	57	echo " --vm-sar-enable Measure CPU utilization in guest VMs using sar."
	58	echo " --host-sar-enable Measure CPU utilization on host using sar."
	59	echo " --sar-delay=INT Wait for X seconds before starting SAR measurement. Default: 0."
	60	echo " --sar-interval=INT Interval (seconds) argument for SAR. Default: 1s."
	61	echo " --sar-count=INT Count argument for SAR. Default: 10."
9f95a23c	62	echo " --vm-throttle-iops=INT I/Os throttle rate in IOPS for each device on the VMs."
11fdf7f2 TL	63	echo " --ctrl-type=TYPE Controller type to use for test:"
	64	echo " spdk_vhost_scsi - use spdk vhost scsi"
	65	echo " spdk_vhost_blk - use spdk vhost block"
9f95a23c	66	echo " kernel_vhost - use kernel vhost scsi"
11fdf7f2	67	echo " Default: spdk_vhost_scsi"
f67539c2	68	echo " --packed-ring Use packed ring support. Requires Qemu 4.2.0 or greater. Default: disabled."
11fdf7f2	69	echo " --use-split Use split vbdevs instead of Logical Volumes"
9f95a23c TL	70	echo " --limit-kernel-vhost=INT Limit kernel vhost to run only on a number of CPU cores."
9f95a23c TL	71	echo " --run-precondition Precondition lvols after creating. Default: true."
f67539c2	72	echo " --precond-fio-bin FIO binary used for SPDK fio plugin precondition. Default: $CONFIG_FIO_SOURCE_DIR/fio."
11fdf7f2 TL	73	echo " --custom-cpu-cfg=PATH Custom CPU config for test."
11fdf7f2 TL	74	echo " Default: spdk/test/vhost/common/autotest.config"
f67539c2 TL	75	echo " --disk-map Disk map for given test. Specify which disks to use, their SPDK name,"
	76	echo " how many times to split them and which VMs should be attached to created bdevs."
	77	echo " Example:"
	78	echo " NVME PCI BDF,Spdk Bdev Name,Split Count,VM List"
	79	echo " 0000:1a:00.0,Nvme0,2,0 1"
	80	echo " 0000:1b:00.0,Nvme1,2,2 3"
11fdf7f2 TL	81	echo "-x set -x for script debug"
	82	exit 0
	83	}
	84
f67539c2	85	function cleanup_lvol_cfg() {
11fdf7f2 TL	86	notice "Removing lvol bdevs"
11fdf7f2 TL	87	for lvol_bdev in "${lvol_bdevs[@]}"; do
f67539c2	88	$rpc_py bdev_lvol_delete $lvol_bdev
11fdf7f2 TL	89	notice "lvol bdev $lvol_bdev removed"
	90	done
	91
	92	notice "Removing lvol stores"
	93	for lvol_store in "${lvol_stores[@]}"; do
f67539c2	94	$rpc_py bdev_lvol_delete_lvstore -u $lvol_store
11fdf7f2 TL	95	notice "lvol store $lvol_store removed"
	96	done
	97	}
	98
f67539c2	99	function cleanup_split_cfg() {
11fdf7f2	100	notice "Removing split vbdevs"
f67539c2 TL	101	for disk in "${disk_cfg_spdk_names[@]}"; do
f67539c2 TL	102	$rpc_py bdev_split_delete ${disk}n1
11fdf7f2 TL	103	done
	104	}
	105
f67539c2 TL	106	function cleanup_parted_config() {
	107	notice "Removing parted disk configuration"
	108	for disk in "${disk_cfg_kernel_names[@]}"; do
	109	parted -s /dev/${disk}n1 rm 1
9f95a23c TL	110	done
	111	}
	112
f67539c2	113	function cleanup_kernel_vhost() {
9f95a23c TL	114	notice "Cleaning kernel vhost configration"
	115	targetcli clearconfig confirm=True
	116	cleanup_parted_config
	117	}
	118
f67539c2 TL	119	function create_vm() {
	120	vm_num=$1
	121	setup_cmd="vm_setup --disk-type=$ctrl_type --force=$vm_num --memory=$vm_memory --os=$VM_IMAGE"
	122	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
	123	x=$(printf %03d $vm_num)
	124	setup_cmd+=" --disks=${wwpn_prefix}${x}"
	125	else
	126	setup_cmd+=" --disks=0"
	127	fi
	128
	129	if $packed_ring; then
	130	setup_cmd+=" --packed"
	131	fi
	132
	133	$setup_cmd
	134	used_vms+=" $vm_num"
	135	echo "Added to used vms"
	136	echo $used_vms
	137	}
	138
	139	function create_spdk_controller() {
	140	vm_num=$1
	141	bdev=$2
	142
	143	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
	144	$rpc_py vhost_create_scsi_controller naa.0.$vm_num
	145	notice "Created vhost scsi controller naa.0.$vm_num"
	146	$rpc_py vhost_scsi_controller_add_target naa.0.$vm_num 0 $bdev
	147	notice "Added LUN 0/$bdev to controller naa.0.$vm_num"
	148	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
	149	if $packed_ring; then
	150	p_opt="-p"
	151	fi
	152
	153	$rpc_py vhost_create_blk_controller naa.0.$vm_num $bdev $p_opt
	154	notice "Created vhost blk controller naa.0.$vm_num $bdev"
	155	fi
	156	}
	157
11fdf7f2 TL	158	while getopts 'xh-:' optchar; do
	159	case "$optchar" in
	160	-)
f67539c2 TL	161	case "$OPTARG" in
	162	help) usage $0 ;;
	163	fio-bin=) fio_bin="--fio-bin=${OPTARG#=}" ;;
	164	fio-jobs=) fio_jobs="${OPTARG#=}" ;;
	165	fio-iterations=) fio_iterations="${OPTARG#=}" ;;
	166	fio-gtod-reduce) fio_gtod="--gtod-reduce" ;;
	167	vm-memory=) vm_memory="${OPTARG#=}" ;;
	168	vm-image=) VM_IMAGE="${OPTARG#=}" ;;
	169	vm-sar-enable) vm_sar_enable=true ;;
	170	host-sar-enable) host_sar_enable=true ;;
	171	sar-delay=) sar_delay="${OPTARG#=}" ;;
	172	sar-interval=) sar_interval="${OPTARG#=}" ;;
	173	sar-count=) sar_count="${OPTARG#=}" ;;
	174	vm-throttle-iops=) vm_throttle="${OPTARG#=}" ;;
	175	ctrl-type=) ctrl_type="${OPTARG#=}" ;;
	176	packed-ring) packed_ring=true ;;
	177	use-split) use_split=true ;;
	178	run-precondition) run_precondition=true ;;
	179	precond-fio-bin=) precond_fio_bin="${OPTARG#=}" ;;
	180	limit-kernel-vhost=) kernel_cpus="${OPTARG#=}" ;;
	181	custom-cpu-cfg=) custom_cpu_cfg="${OPTARG#=}" ;;
	182	disk-map=) disk_map="${OPTARG#=}" ;;
	183	*) usage $0 "Invalid argument '$OPTARG'" ;;
	184	esac
	185	;;
	186	h) usage $0 ;;
	187	x)
	188	set -x
	189	x="-x"
	190	;;
	191	*) usage $0 "Invalid argument '$OPTARG'" ;;
11fdf7f2 TL	192	esac
	193	done
	194
f67539c2	195	rpc_py="$rootdir/scripts/rpc.py -s $(get_vhost_dir 0)/rpc.sock"
11fdf7f2 TL	196
	197	if [[ -n $custom_cpu_cfg ]]; then
	198	source $custom_cpu_cfg
f67539c2 TL	199	vhost_reactor_mask="vhost_${vhost_num}_reactor_mask"
	200	vhost_reactor_mask="${!vhost_reactor_mask}"
	201	vhost_master_core="vhost_${vhost_num}_master_core"
	202	vhost_master_core="${!vhost_master_core}"
11fdf7f2 TL	203	fi
11fdf7f2 TL	204
f67539c2 TL	205	if [[ -z $fio_jobs ]]; then
f67539c2 TL	206	error "No FIO job specified!"
11fdf7f2 TL	207	fi
	208
	209	trap 'error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
11fdf7f2	210
f67539c2 TL	211	if [[ -z $disk_map ]]; then
f67539c2 TL	212	fail "No disk map provided for test. Exiting."
9f95a23c	213	fi
9f95a23c TL	214
	215	# ===== Precondition NVMes if specified =====
	216	if [[ $run_precondition == true ]]; then
	217	# Using the same precondition routine possible for lvols thanks
	218	# to --clear-method option. Lvols should not UNMAP on creation.
f67539c2 TL	219	json_cfg=$rootdir/nvme.json
	220	cat <<- JSON > "$json_cfg"
	221	{"subsystems":[
	222	$("$rootdir/scripts/gen_nvme.sh" --json)
	223	]}
	224	JSON
	225	mapfile -t nvmes < <(grep -oP "Nvme\d+" "$json_cfg")
	226	fio_filename=$(printf ":%sn1" "${nvmes[@]}")
	227	fio_filename=${fio_filename:1}
	228	$precond_fio_bin --name="precondition" \
	229	--ioengine="${rootdir}/build/fio/spdk_bdev" \
	230	--rw="write" --spdk_json_conf="$json_cfg" --thread="1" \
	231	--group_reporting --direct="1" --size="100%" --loops="2" --bs="256k" \
	232	--iodepth=32 --filename="${fio_filename}" \|\| true
9f95a23c TL	233	fi
9f95a23c TL	234
f67539c2 TL	235	set +x
	236	readarray disk_cfg < $disk_map
	237	for line in "${disk_cfg[@]}"; do
	238	echo $line
	239	IFS=","
	240	s=($line)
	241	disk_cfg_bdfs+=(${s[0]})
	242	disk_cfg_spdk_names+=(${s[1]})
	243	disk_cfg_splits+=(${s[2]})
	244	disk_cfg_vms+=("${s[3]}")
	245
	246	# Find kernel nvme names
	247	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
	248	tmp=$(find /sys/devices/pci* -name ${s[0]} -print0 \| xargs sh -c 'ls $0/nvme')
	249	disk_cfg_kernel_names+=($tmp)
	250	IFS=" "
	251	fi
	252	done
	253	unset IFS
	254	set -x
	255
9f95a23c	256	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
f67539c2	257	notice "Configuring kernel vhost..."
9f95a23c	258	trap 'vm_kill_all; sleep 1; cleanup_kernel_vhost; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
f67539c2	259
9f95a23c TL	260	# Split disks using parted for kernel vhost
9f95a23c TL	261	newline=$'\n'
9f95a23c	262	backstores=()
f67539c2 TL	263	for ((i = 0; i < ${#disk_cfg_kernel_names[@]}; i++)); do
	264	nvme=${disk_cfg_kernel_names[$i]}
	265	splits=${disk_cfg_splits[$i]}
	266	notice " Creating extended partition on disk /dev/${nvme}n1"
	267	parted -s /dev/${nvme}n1 mklabel msdos
	268	parted -s /dev/${nvme}n1 mkpart extended 2048s 100%
	269
	270	part_size=$((100 / ${disk_cfg_splits[$i]})) # Split 100% of disk into roughly even parts
	271	echo " Creating ${splits} partitions of relative disk size ${part_size}"
	272	for p in $(seq 0 $((splits - 1))); do
	273	p_start=$((p * part_size))
	274	p_end=$((p_start + part_size))
	275	parted -s /dev/${nvme}n1 mkpart logical ${p_start}% ${p_end}%
	276	sleep 3
	277	done
9f95a23c	278
f67539c2 TL	279	# Prepare kernel vhost configuration
	280	# Below grep: match only NVMe partitions which are not "Extended" type.
	281	# For example: will match nvme0n1p15 but not nvme0n1p1
	282	partitions=$(find /dev -name "${nvme}n1*" \| sort --version-sort \| grep -P 'p(?!1$)\d+')
	283	# Create block backstores for vhost kernel process
	284	for p in $partitions; do
	285	backstore_name=$(basename $p)
	286	backstores+=("$backstore_name")
	287	targetcli backstores/block create $backstore_name $p
	288	done
	289	partitions=($partitions)
	290
	291	# Create kernel vhost controllers and add LUNs
	292	# Setup VM configurations
	293	vms_to_run=(${disk_cfg_vms[i]})
	294	for ((j = 0; j < ${#vms_to_run[@]}; j++)); do
	295	# WWPN prefix misses 3 characters. Need to complete it
	296	# using block backstore number
	297	x=$(printf %03d ${vms_to_run[$j]})
	298	wwpn="${wwpn_prefix}${x}"
	299	targetcli vhost/ create $wwpn
	300	targetcli vhost/$wwpn/tpg1/luns create /backstores/block/$(basename ${partitions[$j]})
	301	create_vm ${vms_to_run[j]}
	302	sleep 1
	303	done
9f95a23c	304	done
f67539c2	305	targetcli ls
11fdf7f2	306	else
f67539c2 TL	307	notice "Configuring SPDK vhost..."
f67539c2 TL	308	vhost_run "${vhost_num}" "--no-gen-nvme" "-p ${vhost_master_core}" "-m ${vhost_reactor_mask}"
9f95a23c TL	309	notice "..."
	310
	311	if [[ $use_split == true ]]; then
f67539c2	312	notice "Configuring split bdevs configuration..."
9f95a23c	313	trap 'cleanup_split_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
f67539c2 TL	314	for ((i = 0; i < ${#disk_cfg_bdfs[@]}; i++)); do
	315	nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
	316	notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"
	317
	318	splits=$($rpc_py bdev_split_create $nvme_bdev ${disk_cfg_splits[$i]})
	319	splits=($splits)
	320	notice "Created splits: ${splits[*]} on Bdev ${nvme_bdev}"
	321	for s in "${splits[@]}"; do
	322	split_bdevs+=($s)
	323	done
	324
	325	vms_to_run=(${disk_cfg_vms[i]})
	326	for ((j = 0; j < ${#vms_to_run[@]}; j++)); do
	327	notice "Setting up VM ${vms_to_run[j]}"
	328	create_spdk_controller "${vms_to_run[j]}" ${splits[j]}
	329	create_vm ${vms_to_run[j]}
9f95a23c	330	done
f67539c2	331	echo " "
11fdf7f2	332	done
9f95a23c TL	333	bdevs=("${split_bdevs[@]}")
9f95a23c TL	334	else
f67539c2	335	notice "Configuring LVOLs..."
9f95a23c	336	trap 'cleanup_lvol_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
f67539c2 TL	337	for ((i = 0; i < ${#disk_cfg_bdfs[@]}; i++)); do
	338	nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
	339	notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"
	340
	341	ls_guid=$($rpc_py bdev_lvol_create_lvstore $nvme_bdev lvs_$i --clear-method none)
9f95a23c	342	lvol_stores+=("$ls_guid")
f67539c2 TL	343	notice "Created Lvol Store: $ls_guid on Bdev $nvme_bdev"
	344
	345	vms_to_run=(${disk_cfg_vms[i]})
	346	for ((j = 0; j < ${disk_cfg_splits[$i]}; j++)); do
9f95a23c	347	free_mb=$(get_lvs_free_mb "$ls_guid")
f67539c2 TL	348	size=$((free_mb / ((${disk_cfg_splits[$i]} - j))))
f67539c2 TL	349	lb_name=$($rpc_py bdev_lvol_create -u $ls_guid lbd_$j $size --clear-method none)
9f95a23c	350	lvol_bdevs+=("$lb_name")
f67539c2 TL	351	notice "Created LVOL Bdev $lb_name on Lvol Store $ls_guid on Bdev $nvme_bdev"
	352
	353	notice "Setting up VM ${vms_to_run[j]}"
	354	create_spdk_controller "${vms_to_run[j]}" ${lb_name}
	355	create_vm ${vms_to_run[j]}
9f95a23c	356	done
f67539c2	357	echo " "
9f95a23c	358	done
f67539c2	359	$rpc_py bdev_lvol_get_lvstores
9f95a23c	360	fi
f67539c2 TL	361	$rpc_py bdev_get_bdevs
f67539c2 TL	362	$rpc_py vhost_get_controllers
11fdf7f2 TL	363	fi
11fdf7f2 TL	364
11fdf7f2 TL	365	# Start VMs
	366	# Run VMs
	367	vm_run $used_vms
	368	vm_wait_for_boot 300 $used_vms
	369
9f95a23c TL	370	if [[ -n "$kernel_cpus" ]]; then
	371	mkdir -p /sys/fs/cgroup/cpuset/spdk
	372	kernel_mask=$vhost_0_reactor_mask
	373	kernel_mask=${kernel_mask#"["}
	374	kernel_mask=${kernel_mask%"]"}
	375
	376	echo "$kernel_mask" >> /sys/fs/cgroup/cpuset/spdk/cpuset.cpus
	377	echo "0-1" >> /sys/fs/cgroup/cpuset/spdk/cpuset.mems
	378
f67539c2	379	kernel_vhost_pids=$(pgrep "vhost" -U root)
9f95a23c TL	380	for kpid in $kernel_vhost_pids; do
	381	echo "Limiting kernel vhost pid ${kpid}"
	382	echo "${kpid}" >> /sys/fs/cgroup/cpuset/spdk/tasks
	383	done
	384	fi
	385
11fdf7f2 TL	386	# Run FIO
	387	fio_disks=""
	388	for vm_num in $used_vms; do
11fdf7f2	389	host_name="VM-$vm_num"
f67539c2	390	vm_exec $vm_num "hostname $host_name"
11fdf7f2 TL	391	vm_start_fio_server $fio_bin $vm_num
	392
	393	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
	394	vm_check_scsi_location $vm_num
	395	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
	396	vm_check_blk_location $vm_num
9f95a23c TL	397	elif [[ "$ctrl_type" == "kernel_vhost" ]]; then
	398	vm_check_scsi_location $vm_num
	399	fi
	400
	401	if [[ -n "$vm_throttle" ]]; then
	402	block=$(printf '%s' $SCSI_DISK)
f67539c2 TL	403	major_minor=$(vm_exec "$vm_num" "cat /sys/block/$block/dev")
	404	vm_exec "$vm_num" "echo \"$major_minor $vm_throttle\" > /sys/fs/cgroup/blkio/blkio.throttle.read_iops_device"
	405	vm_exec "$vm_num" "echo \"$major_minor $vm_throttle\" > /sys/fs/cgroup/blkio/blkio.throttle.write_iops_device"
11fdf7f2 TL	406	fi
	407
	408	fio_disks+=" --vm=${vm_num}$(printf ':/dev/%s' $SCSI_DISK)"
	409	done
	410
	411	# Run FIO traffic
f67539c2 TL	412	for fio_job in ${fio_jobs//,/ }; do
	413	fio_job_fname=$(basename $fio_job)
	414	fio_log_fname="${fio_job_fname%%.*}.log"
	415	for i in $(seq 1 $fio_iterations); do
	416	echo "Running FIO iteration $i for $fio_job_fname"
	417	run_fio $fio_bin --hide-results --job-file="$fio_job" --out="$VHOST_DIR/fio_results" --json $fio_disks $fio_gtod &
	418	fio_pid=$!
	419
	420	if $host_sar_enable \|\| $vm_sar_enable; then
	421	pids=""
	422	mkdir -p $VHOST_DIR/fio_results/sar_stats
	423	sleep $sar_delay
	424	fi
	425
	426	if $host_sar_enable; then
	427	sar -P ALL $sar_interval $sar_count > "$VHOST_DIR/fio_results/sar_stats/sar_stats_host.txt" &
9f95a23c	428	pids+=" $!"
f67539c2 TL	429	fi
	430
	431	if $vm_sar_enable; then
	432	for vm_num in $used_vms; do
	433	vm_exec "$vm_num" "mkdir -p /root/sar; sar -P ALL $sar_interval $sar_count >> /root/sar/sar_stats_VM${vm_num}_run${i}.txt" &
	434	pids+=" $!"
	435	done
	436	fi
	437
9f95a23c TL	438	for j in $pids; do
	439	wait $j
	440	done
9f95a23c	441
f67539c2 TL	442	if $vm_sar_enable; then
	443	for vm_num in $used_vms; do
	444	vm_scp "$vm_num" "root@127.0.0.1:/root/sar/sar_stats_VM${vm_num}_run${i}.txt" "$VHOST_DIR/fio_results/sar_stats"
	445	done
	446	fi
	447
	448	wait $fio_pid
	449	mv $VHOST_DIR/fio_results/$fio_log_fname $VHOST_DIR/fio_results/$fio_log_fname.$i
	450	sleep 1
	451	done
	452
	453	parse_fio_results "$VHOST_DIR/fio_results" "$fio_log_fname"
9f95a23c	454	done
11fdf7f2 TL	455
	456	notice "Shutting down virtual machines..."
	457	vm_shutdown_all
	458
9f95a23c TL	459	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
9f95a23c TL	460	cleanup_kernel_vhost \|\| true
11fdf7f2	461	else
9f95a23c TL	462	notice "Shutting down SPDK vhost app..."
	463	if [[ $use_split == true ]]; then
	464	cleanup_split_cfg
	465	else
	466	cleanup_lvol_cfg
	467	fi
f67539c2	468	vhost_kill "${vhost_num}"
9f95a23c TL	469	fi
	470
	471	if [[ -n "$kernel_cpus" ]]; then
	472	rmdir /sys/fs/cgroup/cpuset/spdk
11fdf7f2	473	fi